Google launched the Gemini Decisions API and Gemini Decision 1 on November 11, 2026, giving developers a model for yes/no checks, fixed choices and rubric scoring across multimodal input. Available through the Gemini API in Google AI Studio, it offers research-tool developers a way to return bounded judgments with confidence estimates.
The launch appears in Google’s Gemini API release notes. The company’s announcement identifies policy checks, tool selection, routing and context grading as intended uses. Support in Gemini in Google Cloud is described as coming soon.
What Gemini Decision 1 returns
According to Google’s model documentation, gemini-decision-1 accepts text, images, audio, video and PDFs. It scores supplied alternatives without generating answer text. The three question types return different values:
- Yes/no: The noul type returns a probability of yes between 0.01 and 0.99, plus confidence.
- Choice: A question evaluates 2 to 255 options and returns the selected option, probabilities for each option and confidence.
- Score: A rubric has 2 to 255 ordered levels. Its result is a probability-weighted expected score, with per-level probabilities and confidence, rather than necessarily an integer grade.
The distinction matters for research evaluation: a score summarising a distribution across rubric levels does not by itself specify which workflow action to take.
A research workflow example: checking, routing and scoring
Illustrative composite: A university repository team could submit a dataset description and supporting PDF, ask whether required documentation is present, choose a review queue, and score documentation completeness. These are potential applications, not a reported university deployment.
Google’s Decisions API guide says questions run independently against shared input and cannot see one another. The application’s code must therefore enforce the relationship between a prerequisite check and a proposed route. A routing answer should not override a failed documentation check merely because both came back in one response.
For an illustrative three-level rubric indexed 0, 1 and 2, suppose the returned probabilities were 0.20, 0.30 and 0.50. The expected score would be 0 × 0.20 + 1 × 0.30 + 2 × 0.50 = 1.30. Level 2 would be the most likely individual level, with confidence 0.50. These demonstration values show why a continuous score and a selected category need separate interpretation.
For a research office documenting such an evaluation, an AI evaluation card can record the rubric and escalation criteria alongside the project’s model evaluation suite.
Confidence is an estimate, not a correctness guarantee
The confidence documentation defines confidence as an estimated probability from 0 to 1 that the top answer is right. For yes/no questions it uses the larger of the yes and no probabilities. High confidence can therefore accompany a negative answer.
Google tells developers to set thresholds using labelled examples from their own traffic and weigh false positives against false negatives. Its guide also warns that choice questions allocate probability even when no option fits, suggesting an explicit catch-all option or a separate scope check. For universities, the implication is that a confidence threshold needs evaluation against the actual task; the number alone cannot establish correctness or institutional approval. CASRAI’s explanation of calibration curves in machine learning provides context for evaluating probability estimates.
Google reports end-to-end latency of 150–250 milliseconds in tests with contexts up to 16K tokens, nearly flat across 1 to 100 questions, and claims the lowest calibration error across its benchmark suite. Those are vendor-reported results; they do not establish the same latency or accuracy for a university’s full workload.
Request limits and availability
The request limits allow 1 to 1,000 questions within a 262,144-token context window covering both shared input and all questions. The model documentation lists Batch API, caching and function calling as unsupported. The guide specifies the Gemini Developer API’s v1beta version.
Google’s rate-limit table lists a free-tier allowance of 200 requests per minute, 50,000 input tokens per minute and 10,000 requests per day. Quotas apply per project, and Google says published rate limits are not guaranteed. A large permitted request can therefore exceed a project’s token-per-minute allowance even while remaining within the model’s context window.
Sources
- Google: Gemini Decisions API: Fast, calibrated multimodal decisions. Checked 2026-11-14.
- Google Gemini API release notes, November 11, 2026 entry. Checked 2026-11-14.
- Google: Gemini Decision 1 model documentation. Checked 2026-11-14.
- Google: Decisions API documentation. Checked 2026-11-14.
- Google: Gemini API rate limits. Checked 2026-11-14.
